Insights to the characterization of cell motility and intercellular communication through a bioimage analysis perspective

Abstract

The study of cell migration is critical for understanding cancer cell biology. Shedding light on the mechanisms that drive cancer cells through metastasis is essential for new treatment development. Although there exist a large variety of processes involved in cell migration such as cell differentiation, tissue stiffening, or intravasation, in this thesis we contribute image analysis methods and statistical tools for the study of two of them: cell motility and inter-cellular communication. In the tumor Extracellular Matrix (ECM) stroma, cells follow a 3D mesenchymal migration mode driven by cellular dendritic protrusions. Together with cellular adhesion to the ECM, cells use elongated protrusions to exert forces and make contractile displacements. Therefore, during cell migration, it is possible to detect varying mechanical patterns which are closely related to the invasive behavior of the cells. However, the relationship between cellular protrusions morphology and dynamics, and cell motility remains largely unknown. Here we propose to analyze time-lapse microscopy videos of cells migrating in 3D Collagen type I matrices. For that, this thesis introduces different image processing approaches to automatically segment cells, and detect their protrusions. Specifically, we develop Deep Learning (DL)-based workflows to accurately segment cells in the microscopy videos. We show the need of combining classical image processing techniques to accurately quantify protrusions in the images. Hence, we assess an additional step to analyze the morphological information of the cell and detect its protrusions’ tips. The information extracted from the image processing enables the study of cell motility and cell protrusion dynamics. We find some ambiguities and limitations in the topological definition of the protrusions seen in the videos, which lead us to redefine such structures to enable its analysis. Preliminary results indicate a relationship between the number and length of cellular protrusions and the cell dynamics. This preliminary characterization of cell protrusions morphology in 3D cancer cell migration can spur researchers to formulate further hypotheses about cell motility and to design more specific biological experiments. Thanks to the automatic quantification of cellular shape and their protrusions in timelapse microscopy videos, it is possible to analyze large datasets for different biological experiments. Nonetheless, one of the outcomes when applying our image processing workflows to biological research is the impracticability of Null-Hypothesis Statistical Tests (NHST) due to the large size of the information extracted (> 1000 cells). Hence, here we also contribute an alternative statistical method towards the analysis of such large datasets. It assesses the differences between compared groups of biological experiments by modeling the p-value of NHST as a function of the sample size and measuring its decay. The results of this method are proven to be robust through simulations and real experimental datasets, among the ones we find the results of the image processing workflow when analyzing the morphology of cells treated with Taxol (a chemotherapy drug). Small (30 − 200 nm) extracellular vesicles (sEVs) are cell-derived nanoscale particles involved in inter-cellular communication. They transport molecular information that goes from the parent to the receiver cell. SEVs are known to be present in a plethora of physiological and pathological processes. In cancer, it is known that they contribute to the changes in the tumor micro-environment or are involved in the formation of the pre-metastatic niche. However, the study of sEVs is a relatively new topic in science so their role in many biological processes is still unknown. Because sEVs are crucial in cellular communication, there is a growing interest on characterizing them and discovering their potential on clinical applications and therapies. Transmission Electron Microscopy (TEM) is the most extended image acquisition technique for the study of nano-scale structures. However, the complexity of nano-scale sEVs sample preparation for TEM prevents the acquisition of clean images. Hence, life-scientists cannot automatically anlyze their images with common user-friendly image processing software. We propose the implementation of a DL-based pipeline for the instance segmentation of sEVs in highly heterogeneous TEM images. The method uses a fully residual U-Net to segment the TEM images and the Radon transform to solve clustered sEVs. We evaluate the approach on three different datasets and show an improved performance over the compared state-of-the-art methods. Despite the potential of all the image processing methods shown here for cancer research, their use still relies on human engineering. The last contribution of this thesis faces the problem of building easy-to-use and open-source environments for the deployment of DL models: deepImageJ. It is a user-friendly plugin for ImageJ to run trained DL models on images in one click. It defines a general model format independent of the neural network architecture or the programming language used to implement it. Thus, it sets a bridge to connect model developers with final users. DeepImageJ interface is designed for bioimage applications: the technicalities of the model are hidden while those relative to the bioimage metadata are summarized. The deepImageJ environment is thought to improve user experience and to support developers’ work. Trained models can be disseminated through the model repository synchronized with the BioImage Model Zoo (https://bioimage.io/). Hence, developers can avoid working on additional scripts to release their pipelines and their work will gain larger visibility. We expect deepImageJ to become a key contributor to the democratization of DL in bioimage analysis.